In the recent week with Claude Fable 5, instead of just using it, I had it run side-by-side comparisons vs Opus on all my real work — same task to both, then a decision-level diff.
Goal: capture its JUDGMENT, not its answers, in a form Opus can load.
What I learned 🧵
How to copy this:
• Distill the diffs into a ~900-word skill file — only rules that changed decisions
• Never let a model describe itself (flattering fiction)
• Demote rules that stop differentiating
• Keep score both ways — its falsified hypotheses are in my record too
Hey @Google, your phone support staff really want to help, but keep giving uncoordinated or bad information. i.e. after authorizing overnight, then failed, asking customers to call @FedEx to make it faster isn't allowed. Also email saying call, then calling says email? #ucandobtr
Hey @bing ... the VERY next thing you have to improve with GPT, is spellcheck on phones. GPT would never write any of the 'corrected' words, ever. I spend more time correcting the corrections than writing. Also #itsneverduck #bingai #gpt4
Anyone else looking forward to being on yet another-another waitlist for AI tomorrow? #OpenAI #ChatGPTPlus #Bing #BingChatGPT #BardAI #AzureOpenAI
Especially for developers, I wrote a handy tool.
GPT-3 backed, this converts, comments, rewrites, writes, and explains code, and more. Lets you use your own OpenAI API key.
crosscodex.com/
2/3 ChatGPT: "In a scenario in which every strategy in the game leads to mutual destruction and the goal is to avoid being destroyed, it may not be possible to logically achieve this goal. In this case, the only way to avoid being destroyed would be to not play the game at all."